POLY02 · 100 TRIALS
4.477 nm
VPEvolve final maximum EPE · Frozen: 5.267 nm
Loading measured curve…
A VIRTUAL PROCESS ENGINEER
Better recipes.
Sharper patterns.
Experience retained.
A virtual process engineer reads process manuals and layout evidence, edits the recipe, and learns from commercial-tool feedback. An LLM reflector and curator turn every trial into guidance for the next.
Dense Metal44 · verified endpoint · measured geometry
70.7%
Poly maximum-EPE reduction
from the initial recipe
28.8%
Metal1 maximum-EPE reduction
from the initial recipe
20/20
Feasible endpoints
across Poly and Metal1
10
Physical trials per window
with AI analysis after each
01 / FROM A RUN TO THE NEXT DECISION
A trial ends.
Its evidence stays.
A lithography recipe combines a short global model setup with thousands of lines of local rules. Engineers add these rules to recognize and repair recurring EPE hotspots and MRC violations.
VPEvolve gives the virtual process engineer process knowledge, layout analysis, recipe editing, and commercial-tool evaluation. An LLM reflector interprets each trial; an LLM curator distills its lessons into a Skill Bank for later decisions.
FIGURE 1 · RECIPE DEVELOPMENTView full size ↗
01Propose a changeGlobal setting or local rule
→02Measure its effectEPE · PVB · MRC
→03Retain the evidenceImprovements and failures
↺04Inform the next trialRetrieve, verify, apply
02 / THE HARNESS
Search the recipe.
Then locate the bottleneck.
The virtual process engineer combines process manuals, layout analysis, recipe editing, and commercial-tool evaluation. After every trial, an LLM reflector analyzes the result and an LLM curator updates evidence-linked skills for the actor to retrieve.
FIGURE 3 · METHODView full size ↗
RECIPE-WIDE INTERVENTION
Tune settings across the layout.
The actor reads the recipe, process manuals, and measured layout feedback before proposing a global parameter edit. The commercial tool evaluates the candidate under fixed quality limits.
recipe & layout→global edit→physical evaluation
Global edits share one ten-trial budget with local rules and diagnostic experiments.
FIXED RETENTION LIMITSMax EPE firstMean EPE ≤ 102% of R0PVB ≤ 102% of R0MRC / SRAF = 0
03 / THE HOTSPOT, UP CLOSE
Global settings.
Local precision.
A measured hotspot remains after global exploration. A pattern-guided rule moves selected mask edges. Inspect the same location before and after local correction—at its true physical scale.
Loading measured hotspot…
IC targetPrinted contourPVBMeasured error segment
The fixed-focus error tracks one coordinate; the full-core maximum searches every gauge. These close-ups illustrate the correction mechanism. The paired experiment below uses ten VPEvolve cases.
SAME START. DIFFERENT NEXT STEP.
Where should the next four evaluations go?
Ten matched cases continue from the same pre-local recipe, with four more trials per branch.
| Layer | Shared start | Global only | Global + local | Added reduction |
|---|
THE MEASURED EFFECT
Allowing local rules reaches 18.012 nm on Metal1 versus 18.583 nm with global edits alone. On Poly, the endpoints are 5.185 and 5.238 nm.
Each branch adds four trials and updates experience. The five Poly and five Metal1 pairs share their starting recipe, feedback, and quality limits; all 20 continuations complete.
04 / ONLINE ACTOR–REFLECTOR–CURATOR RUN
The recipe improves.
The Skill Bank grows.
On 20 windows cropped from FreePDK45-generated full-chip layouts, the actor runs ten physical trials per window. LLM reflection and curation turn each measured result into evidence-linked guidance for the next decision.
Loading measured results…
| Layer | Max EPER0 → final · nm | Mean EPER0 → final · nm | PVBR0 → final · 10⁻³ µm | Success | OPC / LLMtotal calls |
|---|---|---|---|---|---|
| Loading measured results… | |||||
All 20 retained endpoints improve maximum EPE by at least 0.1 nm while meeting the mean-EPE, PVB, MRC, and SRAF-printability limits. Each layer has 110 reported commercial-tool calls: ten trials and one endpoint repeat per case, excluding R0. The actor, reflector, and curator use the same frozen Qwen3.6-27B model.
INSIDE THE PHYSICAL RESULT
One pattern.
Five layers of evidence.
Look through the target layout, corrected mask, assist features, nominal resist contour, and process variation band. Explore two measured dense windows. Switch from the whole layout to a 180 nm detail of the measured hotspot.
Loading measured geometry…
Watch the printed shape change.
Replay every measured trial ↗Dense Poly19 and Metal44 are selected measured recipe cases. Close-ups stay centered on the worst gauge just before local correction begins. The full-window maximum can move elsewhere. All layers are co-registered; resist is the simulated nominal contour, not a photograph.
05 / MATCHED BASELINES
Seven baselines.
One recipe budget.
VPEvolve, six agent baselines, and Bayesian optimization work on the same 20 Poly and Metal1 windows. Each receives ten recipe trials per case, with the same physical evaluation and quality limits.
MEAN WINDOW-MAX EPELoading…
Comparing matched endpoints
| Method | Max EPE ↓nm | Mean EPE ↓nm | PVB ↓10⁻³ µm | Complete | Success | OPC / LLMtotal calls |
|---|---|---|---|---|---|---|
| Loading final results… | ||||||
Values are means of retained recipes for the selected group; use the layer tabs for separate Poly and Metal1 results. Success requires a feasible maximum-EPE reduction of at least 0.1 nm. Each method uses ten candidate trials per case; reported OPC calls exclude R0 and include endpoint repeats.
06 / EXPERIENCE EVOLUTION
Every trial can sharpen
the next decision.
Full VPEvolve and four mechanism variants run on the same 20 cases. The comparison tests accumulated experience, domain response reflection, scoped revision, and experiment guidance.
Loading final mechanism results…
WHAT THE ABLATION SHOWS
Loading measured comparison…
Cards report mean maximum EPE for the best feasible recipes retained across each group. Completion counts distinguish runs that finished all ten trials and passed endpoint repeat.
PAIRED EXPERIENCE STUDY
Same recipe. Different experience.
At the fifth trial, each case branches from the same retained recipe and raw trial history. One branch receives ACE experience; the other receives VPEvolve experience. Both snapshots stay fixed during five additional trials.
| Layer | Shared start | ACE experience | VPEvolve experience | VPEvolve lower by |
|---|---|---|---|---|
| Loading paired results… | ||||
Ten paired cases per layer. Both branches use the same observations, model, available actions, and remaining evaluation budget; only the supplied experience differs.
07 / SIX-MODEL COMPARISON
The full loop works
across models.
Six language models each serve as actor, reflector, and curator on the same benchmark and ten-trial budget.
FIGURE 7 · CROSS-MODEL RESULTSView full size ↗
| Layer | Model | Max EPE ↓nm | Mean EPE ↓nm | PVB ↓10⁻³ µm |
|---|---|---|---|---|
| Loading cross-model results… | ||||
Each model fills all three roles in the complete VPEvolve loop. Starting recipes, cases, actions, and physical trial budgets are matched.
08 / LONG-HORIZON OPTIMIZATION
Keep improving
beyond ten trials.
On Poly02 and Metal29, Full VPEvolve and Frozen Experience start from the same recipe and run toward a 100-trial horizon. Follow the retained recipe after every measured trial.
METAL29 · EXTENDED SEARCH
10.912 nm
VPEvolve at trial 100 · Frozen: 13.348 nm at trial 57
Loading measured curve…
Poly runs complete 100 trials. Metal29 Frozen stops after trial 57; at that common point VPEvolve is at 12.470 nm versus 13.348 nm for Frozen. VPEvolve continues to trial 100. Curves end at the last observed trial and omit endpoint repeats.
09 / OPEN CODE & DATA
Inspect the agent.
Run the optics.
The VPEvolve code, a standalone differentiable lithography simulator, and 20 GDS windows from FreePDK45-generated full-chip layouts are available in a dedicated Apache-2.0 repository.
VPEvolve code
Agent and harness, experiment configurations, initial recipes, the simulator, and all 20 layouts.
Download full source ↓Lithography simulator
Differentiable optical and resist simulation, analytic modes, examples, tests, and the layout set.
Download simulator ↓Layouts and manifest
Ten Poly and ten Metal1 windows, with layer labels, scoring-core coordinates, and SHA-256 hashes.
Download layouts ↓The agent pipeline uses a separately licensed commercial tool and process model. The standalone simulator runs optical and resist experiments. Browse the repository ↗